We can take the autoencoder architecture further by forcing it to learn more important features about the input data. By adding noise to the input images and having the original ones as the target, the model will try to remove this noise and learn important features about them in order to come up with meaningful reconstructed images in the output. This kind of CAE architecture can be used to remove noise from input images. This specific variation of autoencoders is called denoising autoencoder:
Figure 10: Examples of original images and the same images after adding a bit of Gaussian noise
So let's start off by implementing the architecture in the following figure. The only extra thing that we have added to this denoising autoencoder architecture is some noise in the original input image:
Figure 11: General denoising architecture of autoencoders